More doesn't mean better.
Uber has achieved near-total internal AI adoption. About 95% of its engineers use AI monthly and around 70% of committed code is AI-generated. Not only that, they burned through their 2026 token budget in just four months.
But there is an ugly truth behind those numbers. The Uber leadership admits they can’t clearly link this surge in token consumption to more or better consumer-facing features. They aren't the only ones.
Uber's COO Andrew Macdonald publicly conceded that despite massive usage, the "link is not there yet” between their AI activity and meaningful product output. Heavy usage of tools like Claude Code cost more, but are simply not delivering more value (yet?).
This is the AI productivity paradox. Faster isn't necessarily better.
Whilst AI coding tools increase individual output (more code, faster tasks), company-level outcomes often stall due to constraints elsewhere (reviews, QA, integration, decision-making). Often human based constraints.
WHY IT MATTERS
The missing link isn’t more AI it’s humans behaving differently.
If teams measure effort (tokens, commits, usage) instead of outcomes (customer value, cycle time, quality), AI accelerates the wrong things faster.
Transformation leaders need to pause, take a moment and rewire incentives, workflows, and decision rights, not just deploy tools. Otherwise, you are building a Ferrari engine into a traffic jam.
This is similar to the capacity constraints Scaled Agile often walks into.
WHAT TO WATCH FOR
→ Measures: watch for orgs shifting from usage metrics to value metrics: deployment frequency, lead time, feature adoption, defect rates.
→ Culturally: are teams rewarded for AI usage or for shipped impact?
→ Structurally: are review, testing, and governance layers being redesigned to match AI-speed development, or becoming the new bottleneck?
LIMITATIONS OF THE REPORTING
The data is still early, anecdotal, and skewed toward engineering metrics rather than end-to-end value creation. AI’s impact may diffuse, lag, or hide inside broader system changes. The signal is probably real, but the causal map is still to be built.
SOURCE
https://www.businessinsider.com/uber-coo-andrew-macdonald-ai-token-spending-harder-justify-2026-5
BESCI AI OPINION
This isn't a surprise to experienced behaviouralists and change experts. Just because the car goes faster doesn't mean that you should, or that you will get there any faster.
The terrain matters. In this case the context and environment that you are working in.
The same constraints that always existed are there, just more visible as the previous ones have been eliminated.
The ones left are the messy, political, siloed humans ones that we navigate daily.
Is this our moment, as experts in changing human behaviour?
I truly hope so.
